Kernel-based Semantic Role Labeling

Automatic semantic parsing has always been one of the main tasks of naturallanguage understanding. Through deep semantic parsing, the natural language can betranslated into the formal language, so that computers can communicate with humanbeings smoothly. To that end, it has been going on for years of effort. However, be-cause this issue is too complex, the results are not resolved ideally until now. Shallowsemantic parsing is a simplified deep semantic parsing. It only labels predicate relatedconstituents with semantic roles in a sentence, such as agent, patient, time, place, andso on. The technique can promote many applications, such as question and answering,information extraction, and machine translation. Semantic role labeling is a way toachieve shallow semantic parsing. It has many advantages, such as clear definitionand easy to evaluate. More and more researchers have paid much attention on it inrecent years.At present, to identify and classify semantic roles, the mainstream studies ofsemantic role labeling focus on the using of a variety of statistical machine learningtechniques and all kinds of linguistics features. In recent years, studies have shownthat machine learning model is not the primary factor to effect the semantic role la-beling performance, but the using of the features. Therefore, in order to improvethe system performance, detailed features engineering work is essential. However, asmore and more features have been added, the interaction among these features hasbecome more and more serious. It makes the growth trend of system performancegradually slowing down and reaching an upper bound. So we must find new ways tosolve this problem.Through the combination or decomposition of existing features, kernel-basedmethods can map a low-dimensional feature space into a higher-dimensional fea-ture space. Thereby, it makes the problem which is not easy to distinguish in low-dimensional feature spaces becoming addressed in high-dimensional feature spaces.We make use of the advantages of this method and apply to the semantic rolelabeling task. In addition to use existing kernel-based methods, we propose a varietyof new methods. At first, we build a baseline semantic role labeling system, which uses a featurevector to represent a classification object and uses a polynomial kernel to combinefeatures automatically. The evaluation results show that, when using a 2nd-order poly-nomial kernel to combine each feature pair, the system can achieve the state-of-the-artperformance, which is based on single syntactic parser.Then, for our baseline system, it has the problem of difficultly representing struc-ture features. We use a convolution tree kernel to decompose these larger structure fea-tures and compute the kernel function in polynomial time. However, the traditionalconvolution tree kernel confuses the different features used in semantic role labeling.Therefore, we provide a hybrid convolution tree kernel to make fusion different con-volution tree kernels, which can model different features with different kernels. Theeventuation results show that the novel method is better than the traditional convo-lution tree kernel. At last, we combine the hybrid convolution tree kernel and the2nd-order polynomial kernel into a composite kernel. The composite kernel outper-forms either of the two individual kernels.However, the standard convolution tree kernel requires exact matching betweentwo sub-trees, without taking into account the similar structures which have the samesemantic roles. Therefore, we propose a new grammar-driven convolution tree kernel.In the design process of the kernel, we integrate linguistic knowledge, and allow thenode and the structure approximate matching. The new kernel outperforms the stan-dard convolution tree kernel. Finally, also combined with the polynomial kernel, thesystem achieves a better performance.At last, we use the methods described above to build the state-of-the-art Chinesesemantic role labeling system. The main contribution is that we propose more newChinese oriented features. On the other hand, we first use the kernel-based methodsin Chinese semantic role labeling. The final performance trend is consistent with theEnglish one, which also proves that our kernel-based methods are effective.